Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl
view MasterLoop.c @ 111:b1817e2451b1
add task # to counter stats
| author | Nina Engelhardt |
|---|---|
| date | Fri, 05 Aug 2011 18:03:07 +0200 |
| parents | 724c7a0b687f |
| children | fac1d896f6e9 |
line source
1 /*
2 * Copyright 2010 OpenSourceStewardshipFoundation
3 *
4 * Licensed under BSD
5 */
9 #include <stdio.h>
10 #include <stddef.h>
12 #include "VMS.h"
13 #include "ProcrContext.h"
16 //===========================================================================
17 void inline
18 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
19 VirtProcr *masterPr );
21 //===========================================================================
25 /*This code is animated by the virtual Master processor.
26 *
27 *Polls each sched slot exactly once, hands any requests made by a newly
28 * done slave to the "request handler" plug-in function
29 *
30 *Any slots that need a virt procr assigned are given to the "schedule"
31 * plug-in function, which tries to assign a virt procr (slave) to it.
32 *
33 *When all slots needing a processor have been given to the schedule plug-in,
34 * a fraction of the procrs successfully scheduled are put into the
35 * work queue, then a continuation of this function is put in, then the rest
36 * of the virt procrs that were successfully scheduled.
37 *
38 *The first thing the continuation does is busy-wait until the previous
39 * animation completes. This is because an (unlikely) continuation may
40 * sneak through queue before previous continuation is done putting second
41 * part of scheduled slaves in, which is the only race condition.
42 *
43 */
45 /*May 29, 2010 -- birth a Master during init so that first core loop to
46 * start running gets it and does all the stuff for a newly born --
47 * from then on, will be doing continuation, but do suspension self
48 * directly at end of master loop
49 *So VMS__init just births the master virtual processor same way it births
50 * all the others -- then does any extra setup needed and puts it into the
51 * work queue.
52 *However means have to make masterEnv a global static volatile the same way
53 * did with readyToAnimateQ in core loop. -- for performance, put the
54 * jump to the core loop directly in here, and have it directly jump back.
55 *
56 *
57 *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
58 * avoids the suspected bug in the system stack that causes bizarre faults
59 * at random places in the system code.
60 *
61 *So, this function is coupled to each of the MasterVPs, -- meaning this
62 * function can't rely on a particular stack and frame -- each MasterVP that
63 * animates this function has a different one.
64 *
65 *At this point, the masterLoop does not write itself into the queue anymore,
66 * instead, the coreLoop acquires the masterLock when it has nothing to
67 * animate, and then animates its own masterLoop. However, still try to put
68 * several AppVPs into the queue to amortize the startup cost of switching
69 * to the MasterVP. Note, don't have to worry about latency of requests much
70 * because most requests generate work for same core -- only latency issue
71 * is case when other cores starved and one core's requests generate work
72 * for them -- so keep max in queue to 3 or 4..
73 */
74 void masterLoop( void *initData, VirtProcr *animatingPr )
75 {
76 int32 slotIdx, numSlotsFilled;
77 VirtProcr *schedVirtPr;
78 SchedSlot *currSlot, **schedSlots;
79 MasterEnv *masterEnv;
80 VMSQueueStruc *readyToAnimateQ;
82 SlaveScheduler slaveScheduler;
83 RequestHandler requestHandler;
84 void *semanticEnv;
86 int32 thisCoresIdx;
87 VirtProcr *masterPr;
88 volatile VirtProcr *volatileMasterPr;
90 volatileMasterPr = animatingPr;
91 masterPr = (VirtProcr*)volatileMasterPr; //used to force re-define after jmp
93 //First animation of each MasterVP will in turn animate this part
94 // of setup code.. (VP creator sets up the stack as if this function
95 // was called normally, but actually get here by jmp)
96 //So, setup values about stack ptr, jmp pt and all that
97 //masterPr->nextInstrPt = &&masterLoopStartPt;
100 //Note, got rid of writing the stack and frame ptr up here, because
101 // only one
102 // core can ever animate a given MasterVP, so don't need to communicate
103 // new frame and stack ptr to the MasterVP storage before a second
104 // version of that MasterVP can get animated on a different core.
105 //Also got rid of the busy-wait.
108 //masterLoopStartPt:
109 while(1){
111 //============================= MEASUREMENT STUFF ========================
112 #ifdef MEAS__TIME_MASTER
113 //Total Master time includes one coreloop time -- just assume the core
114 // loop time is same for Master as for AppVPs, even though it may be
115 // smaller due to higher predictability of the fixed jmp.
116 saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
117 #endif
118 //========================================================================
120 masterEnv = (MasterEnv*)_VMSMasterEnv;
122 //GCC may optimize so doesn't always re-define from frame-storage
123 masterPr = (VirtProcr*)volatileMasterPr; //just to make sure after jmp
124 thisCoresIdx = masterPr->coreAnimatedBy;
125 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
126 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
128 requestHandler = masterEnv->requestHandler;
129 slaveScheduler = masterEnv->slaveScheduler;
130 semanticEnv = masterEnv->semanticEnv;
133 //Poll each slot's Done flag
134 numSlotsFilled = 0;
135 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
136 {
137 currSlot = schedSlots[ slotIdx ];
139 if( currSlot->workIsDone )
140 {
141 currSlot->workIsDone = FALSE;
142 currSlot->needsProcrAssigned = TRUE;
144 //process requests from slave to master
145 //====================== MEASUREMENT STUFF ===================
146 #ifdef MEAS__TIME_PLUGIN
147 int32 startStamp1, endStamp1;
148 saveLowTimeStampCountInto( startStamp1 );
149 #endif
150 #ifdef MEAS__PERF_COUNTERS
151 int lastRecordIdx = currSlot->procrAssignedToSlot->counter_history_array_info->numInArray -1;
152 CounterRecord* lastRecord = currSlot->procrAssignedToSlot->counter_history[lastRecordIdx];
153 lastRecord->req_core = thisCoresIdx;
154 saveCyclesAndInstrs(thisCoresIdx,lastRecord->req_cycles,lastRecord->req_instrs);
155 //End of task, start of next task
156 //print counters from last run
157 print_record_csv_to_file(lastRecord,_VMSMasterEnv->counteroutput);
158 //create new entry in record array
159 CounterRecord* newRecord = VMS__malloc(sizeof(CounterRecord));
160 newRecord->req_core = thisCoresIdx;
161 newRecord->vp_id = currSlot->procrAssignedToSlot->procrID;
162 newRecord->task_position = lastRecord->task_position + 1;
163 getReturnAddressBeforeLibraryCall(currSlot->procrAssignedToSlot, &(newRecord->addr_of_libcall_for_req));
164 addToDynArray( (void*) newRecord, currSlot->procrAssignedToSlot->counter_history_array_info);
165 lastRecord = newRecord;
166 #endif
167 //============================================================
168 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
169 //====================== MEASUREMENT STUFF ===================
170 #ifdef MEAS__TIME_PLUGIN
171 saveLowTimeStampCountInto( endStamp1 );
172 addIntervalToHist( startStamp1, endStamp1,
173 _VMSMasterEnv->reqHdlrLowTimeHist );
174 addIntervalToHist( startStamp1, endStamp1,
175 _VMSMasterEnv->reqHdlrHighTimeHist );
176 #endif
177 #ifdef MEAS__PERF_COUNTERS
178 //done with constraints check
179 saveCyclesAndInstrs(thisCoresIdx,lastRecord->sc_done_cycles,lastRecord->sc_done_instrs);
180 saveLowTimeStampCountInto(lastRecord->blocked_timestamp);
181 #endif
182 //============================================================
183 }
184 if( currSlot->needsProcrAssigned )
185 { //give slot a new virt procr
186 #ifdef MEAS__PERF_COUNTERS
187 //start assigner
188 uint64 tmp_cycles;
189 uint64 tmp_instrs;
190 saveCyclesAndInstrs(thisCoresIdx,tmp_cycles,tmp_instrs);
191 #endif
192 schedVirtPr =
193 (*slaveScheduler)( semanticEnv, thisCoresIdx );
195 if( schedVirtPr != NULL )
196 { currSlot->procrAssignedToSlot = schedVirtPr;
197 schedVirtPr->schedSlot = currSlot;
198 currSlot->needsProcrAssigned = FALSE;
199 numSlotsFilled += 1;
200 #ifdef MEAS__PERF_COUNTERS
201 //end assigner
202 int lastRecordIdx = currSlot->procrAssignedToSlot->counter_history_array_info->numInArray -1;
203 CounterRecord* lastRecord = currSlot->procrAssignedToSlot->counter_history[lastRecordIdx];
204 lastRecord->assigning_core = thisCoresIdx;
205 lastRecord->start_assign_cycles = tmp_cycles;
206 lastRecord->start_assign_instrs = tmp_instrs;
207 saveCyclesAndInstrs(thisCoresIdx,lastRecord->end_assign_cycles,lastRecord->end_assign_instrs);
208 #endif
209 writeVMSQ( schedVirtPr, readyToAnimateQ );
210 }
211 }
212 }
215 #ifdef USE_WORK_STEALING
216 //If no slots filled, means no more work, look for work to steal.
217 if( numSlotsFilled == 0 )
218 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
219 }
220 #endif
223 #ifdef MEAS__TIME_MASTER
224 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
225 #endif
227 masterSwitchToCoreLoop(animatingPr);
228 flushRegisters();
229 }//MasterLoop
232 }
236 /*This has a race condition -- the coreloops are accessing their own queues
237 * at the same time that this work-stealer on a different core is trying to
238 */
239 void inline
240 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
241 VirtProcr *masterPr )
242 {
243 VirtProcr *stolenPr;
244 int32 coreIdx, i;
245 VMSQueueStruc *currQ;
247 stolenPr = NULL;
248 coreIdx = masterPr->coreAnimatedBy;
249 for( i = 0; i < NUM_CORES -1; i++ )
250 {
251 if( coreIdx >= NUM_CORES -1 )
252 { coreIdx = 0;
253 }
254 else
255 { coreIdx++;
256 }
257 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
258 if( numInVMSQ( currQ ) > 0 )
259 { stolenPr = readVMSQ (currQ );
260 break;
261 }
262 }
264 if( stolenPr != NULL )
265 { currSlot->procrAssignedToSlot = stolenPr;
266 stolenPr->schedSlot = currSlot;
267 currSlot->needsProcrAssigned = FALSE;
269 writeVMSQ( stolenPr, readyToAnimateQ );
270 }
271 }
273 /*This algorithm makes the common case fast. Make the coreloop passive,
274 * and show its progress. Make the stealer control a gate that coreloop
275 * has to pass.
276 *To avoid interference, only one stealer at a time. Use a global
277 * stealer-lock.
278 *
279 *The pattern is based on a gate -- stealer shuts the gate, then monitors
280 * to be sure any already past make it all the way out, before starting.
281 *So, have a "progress" measure just before the gate, then have two after it,
282 * one is in a "waiting room" outside the gate, the other is at the exit.
283 *Then, the stealer first shuts the gate, then checks the progress measure
284 * outside it, then looks to see if the progress measure at the exit is the
285 * same. If yes, it knows the protected area is empty 'cause no other way
286 * to get in and the last to get in also exited.
287 *If the progress measure at the exit is not the same, then the stealer goes
288 * into a loop checking both the waiting-area and the exit progress-measures
289 * until one of them shows the same as the measure outside the gate. Might
290 * as well re-read the measure outside the gate each go around, just to be
291 * sure. It is guaranteed that one of the two will eventually match the one
292 * outside the gate.
293 *
294 *Here's an informal proof of correctness:
295 *The gate can be closed at any point, and have only four cases:
296 * 1) coreloop made it past the gate-closing but not yet past the exit
297 * 2) coreloop made it past the pre-gate progress update but not yet past
298 * the gate,
299 * 3) coreloop is right before the pre-gate update
300 * 4) coreloop is past the exit and far from the pre-gate update.
301 *
302 * Covering the cases in reverse order,
303 * 4) is not a problem -- stealer will read pre-gate progress, see that it
304 * matches exit progress, and the gate is closed, so stealer can proceed.
305 * 3) stealer will read pre-gate progress just after coreloop updates it..
306 * so stealer goes into a loop until the coreloop causes wait-progress
307 * to match pre-gate progress, so then stealer can proceed
308 * 2) same as 3..
309 * 1) stealer reads pre-gate progress, sees that it's different than exit,
310 * so goes into loop until exit matches pre-gate, now it knows coreloop
311 * is not in protected and cannot get back in, so can proceed.
312 *
313 *Implementation for the stealer:
314 *
315 *First, acquire the stealer lock -- only cores with no work to do will
316 * compete to steal, so not a big performance penalty having only one --
317 * will rarely have multiple stealers in a system with plenty of work -- and
318 * in a system with little work, it doesn't matter.
319 *
320 *Note, have single-reader, single-writer pattern for all variables used to
321 * communicate between stealer and victims
322 *
323 *So, scan the queues of the core loops, until find non-empty. Each core
324 * has its own list that it scans. The list goes in order from closest to
325 * furthest core, so it steals first from close cores. Later can add
326 * taking info from the app about overlapping footprints, and scan all the
327 * others then choose work with the most footprint overlap with the contents
328 * of this core's cache.
329 *
330 *Now, have a victim want to take work from. So, shut the gate in that
331 * coreloop, by setting the "gate closed" var on its stack to TRUE.
332 *Then, read the core's pre-gate progress and compare to the core's exit
333 * progress.
334 *If same, can proceed to take work from the coreloop's queue. When done,
335 * write FALSE to gate closed var.
336 *If different, then enter a loop that reads the pre-gate progress, then
337 * compares to exit progress then to wait progress. When one of two
338 * matches, proceed. Take work from the coreloop's queue. When done,
339 * write FALSE to the gate closed var.
340 *
341 */
342 void inline
343 gateProtected_stealWorkInto( SchedSlot *currSlot,
344 VMSQueueStruc *myReadyToAnimateQ,
345 VirtProcr *masterPr )
346 {
347 VirtProcr *stolenPr;
348 int32 coreIdx, i, haveAVictim, gotLock;
349 VMSQueueStruc *victimsQ;
351 volatile GateStruc *vicGate;
352 int32 coreMightBeInProtected;
356 //see if any other cores have work available to steal
357 haveAVictim = FALSE;
358 coreIdx = masterPr->coreAnimatedBy;
359 for( i = 0; i < NUM_CORES -1; i++ )
360 {
361 if( coreIdx >= NUM_CORES -1 )
362 { coreIdx = 0;
363 }
364 else
365 { coreIdx++;
366 }
367 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
368 if( numInVMSQ( victimsQ ) > 0 )
369 { haveAVictim = TRUE;
370 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
371 break;
372 }
373 }
374 if( !haveAVictim ) return; //no work to steal, exit
376 //have a victim core, now get the stealer-lock
377 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
378 UNLOCKED, LOCKED );
379 if( !gotLock ) return; //go back to core loop, which will re-start master
382 //====== Start Gate-protection =======
383 vicGate->gateClosed = TRUE;
384 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
385 while( coreMightBeInProtected )
386 { //wait until sure
387 if( vicGate->preGateProgress == vicGate->waitProgress )
388 coreMightBeInProtected = FALSE;
389 if( vicGate->preGateProgress == vicGate->exitProgress )
390 coreMightBeInProtected = FALSE;
391 }
393 stolenPr = readVMSQ ( victimsQ );
395 vicGate->gateClosed = FALSE;
396 //======= End Gate-protection =======
399 if( stolenPr != NULL ) //victim could have been in protected and taken
400 { currSlot->procrAssignedToSlot = stolenPr;
401 stolenPr->schedSlot = currSlot;
402 currSlot->needsProcrAssigned = FALSE;
404 writeVMSQ( stolenPr, myReadyToAnimateQ );
405 }
407 //unlock the work stealing lock
408 _VMSMasterEnv->workStealingLock = UNLOCKED;
409 }